Ieee Mlsp 09 Luo Based
-
Upload
gheorghe-gardu -
Category
Documents
-
view
221 -
download
0
Transcript of Ieee Mlsp 09 Luo Based
-
8/9/2019 Ieee Mlsp 09 Luo Based
1/6
HYPERSPECTRAL IMAGE
CLASSIFICATION BASED ON
SPECTRAL
AND
GEOMETRICAL FEATURES
Bin Luo and Jocelyn Chanussot
GIPSA-Lab, 961 rue de la Houille Blanche, 38402 Grenoble, France.
G
ABSTRACT
In this paper, we propose to integrate geometrical features,
such as the characteristic scales
of
structures, with spectral
features for the classification of hyperspectral images. The
spectral features which only describe the material
of
struc
tures can not distinguish objects made by the same material
but with different semantic meanings (such as the roofs
of
some buildings and the roads). The use of geometrical fea
tures is therefore necessary. Moreover, since the dimension
of
a hyperspectral image is usually very high, we use linear un
mixing algorithm to extract the endmemebers and their abun
dance maps in order to represent compactly the spectral infor
mation. Afterwards, with the help of these abundance maps,
we propose a method based on topographic map
of
images to
estimate local scales of structures in hyperspectral images.
The experiment shows that the geometrical features can
improve the classification results, especially for the classes
made by the same material but with different semantic mean
ings. When compared to the traditional contextual features
(such as morphological profiles), the local scale feature pro
vides satisfactory results without considerably increasing the
feature dimension.
1. INTRODUCTION
The classification
of
hyperspectral remote sensing images is
a challenging task, since the data dimension is considerable
for traditional classification algorithms, typically several hun
dreds of spectral bands are acquired for each image. These
spectral bands can provide very rich spectral information
of
each pixel in order to identify the material of the objects.
However, the spectral information alone sometimes does not
allow the separation of structures. For example, the roofs of
some buidings and the roads can be made by the same ma
terial (alsphalt). Therefore, contextual information, geomet
rical features for example, is necessary for classification task
of
hyperspectral images.
The major difficulty to calculate the contextual informa
tion
of
hyper spectral images is the high dimension
of
the data.
This work is funded by French ANR project VAHINE.
978-1-4244-4948-4/09/ 25.00 © 2009 IEEE
Toreduce the data dimension, both supervised and non super
vised methods are proposed. The supervisedmethods, such as
band selection [1] [2], Decision Boundary feature extraction
and Non-Weighted feature extraction [3], transform the data
according to the training set in order to improve the separabil
ity
of
the data. However, the supervised methods depend on
the quality
of
training set. The unsupervised methods, such
as PCA (principle component analysis) or ICA (independant
component analysis), optimise some statistical criteria (such
as the most un correlated components or the most independant
components) to project the data onto a sub space with lower
dimension. The application
of
these methods on hyperspec
tral data can be found in [4] [5] [6]. However, the components
obtained by optimising the statistical criteria do not neces
sarily have physical meanings. In this paper, we propose to
use VertexComponent Analysis (VCA) to reduce the dimen
sion of hyperspectral data [7]. We suppose that the spectrum
of
each pixel is a linear mixture
of
the spectra
of
different
chemical species (refered as endmembers). This linear mix
ture model is physically valid for the reflectance
of
the surface
of the Earth without being affected by the aerosol. The VCA
can separate the spectra
of
these endmembers and estimate
their spatial abundances. Since the number
of
the endmem
bers is much less than the number
of
spectral bands, these
abundance maps can be considered as a compact represen
tation
of
spectral information provided by the hyperspectral
image.
In order to extract the contextual information from remote
sensing images, one can find the methods based on Markov
Random Field (MRF) [8]. However, the MRF based meth
ods provide only statistical information on the neighborhood
of the considered pixel. Another family of methods for con
textual information extraction are based on morpholigical op
erators, which allow to extract descriptive features, such as
the geometrical features about the structures [9]. [4] [5] ex
tract the extended morphological profiles (EMP) of the princi
ple components
of
hyperspectral images for the classification.
However, since all morphological methods require a structred
element, the features extracted by such methods depend on
the used structured element. Moreover, EMP increase con-
-
8/9/2019 Ieee Mlsp 09 Luo Based
2/6
Fig.
1.
Scheme of the paper.
siderably the feature dimension. For example, in order to de
scribe the geometrical information of one pixel on a principle
component, the EMP method in [4] [5] uses 8 values. In [10],
based on topographic map of gray scale image, the authors
define a characteristic scale of each pixel in panchromatic re
mote sensing images. The main idea isthat for each pixel, the
contrasted structure containing it is extracted, and the scale
of this structure is defined as the scale of this pixel . In this
paper, we try to extend the algorithm presented in [10] for
extracting a local characteristic scale for each pixel on hyper
spectral imageswith the help
of
the abundancemaps obtained
by VCA. The main advantages
of our proposed method when
compared to EMP are two-fold. At one hand, no structured
element is required for this method. Secondly, we use only
one value (the local scale) to describe the geometrical feature
of
a spatial position rather than the EMP which need many
values.
I= MS
n
2. LINEAR UNMIXING OF HYPERSPECTRAL DATA
We note X the matrix representing the hyperspectral im
age cube , where X = {X I
,X 2
, ,X N
a
}
and Xk =
{ XI
,k
, X2 ,k , . . .
x
,,, d
T
Xl
,k
is the value of the kth pixel at
the lth band. We assume that the spectrum
of
each pixel is a
linear mixture of the spectra of
N;
endmembers, leading to
the following model:
where M
=
{m-, m2,
.. .
,
m
NJ
is the mixing matrix, where
m., denotes the spectral signature of the nth endmember.
S = { 5 I , 5 2 ,
. .
. , 5N o } T is the abundance matrix where
5
n
=
{Sn
, l , S n ,2 , . . . , S n , N
a
} S n
,k
E
[0 ,1] is the abun
dance
of
the nth endmember at the kth pixel). n stands for
the additive noise of the image. In [7] , the Vertex Compo
nent Analysis (VCA) is proposed as an effecient method for
extracting the endmembers which are linearly mixed. The
main idea is to extract the vertex of the simplex formed by
M which contains all the data vectors in X. The sum of the
abundances of different endmembers at each pixel is one, i.e,
L : Sn
,k
= 1, which is called the sum-fa-one condition.
Therefore the data vectors Xl are always inside a simplex of
which the vertex are the spectra
of
the endmembers. VCA it
eratively projects the data onto the direction orthogonal to the
subspace spanned by the already determined endmembers.
And the extreme of this project ion is the new endmember
signature. The algorithm stops when all the
p
endmembers
are extracted, where p is the number of endmembers. Even
though the number of endmembers is much smaller than the
number of spectral bands, i.e,
p
-
8/9/2019 Ieee Mlsp 09 Luo Based
3/6
map [11], which can be obtained by Fast Level Set Trans
formation (FLST) [12], represents an image by an inclusion
tree
of
the shapes (which are defined as the connected com
ponents of level sets) . An example of such inclusion tree is
shown in Figure 2 . For each pixel
x,
y), there is a branch
of
shapes fi(X,y)
f
i
l
C
f i) containing it. Note
l J
i) the
gray level of the shape fi(X,y), S(Ji) its area and
P J
i)
its perimeter. The contrast
of
the shape fi(X,y) is defined
as C Ji) = II Ji+ l ) - I( Ji) The most contrasted shape
Ji(x,y)
of
a given pixel is defined as the shape containing
this pixel,
of
which the contrast is the most important, i.e.
•
(a)
A :5255
E(x,y)
= S J
i(x,y))jP(Ji(x,y)).
3
(5)
D :5128
G
C ,,40
F 2:150
(b)
H :50
E(x,y)
=
Eft(x, y)
E 2;128
Fig. 2. Example
ofFLST:
(a) Synthetic
image;
(b) Inclusion
tree obtained with FLST.
object is mainlymade by the
nth
endmember, the scale values
calculated on the abundance maps of the other endmembers
for this object have no meaning. Therefore, we try to define
one single scale value for each pixel which corresponds to
the scale of this pixel on the most significant abundancemap.
More precisely, since the values of different abundance maps
are comparable, the characteristic scale at the spatial position
x,
y for a hyperspectral image is defined as
8 x ,y =
{ll
x ,y ,
.. .
INc(x ,y ,E(x ,y } .
6
where n
=
argmaxn{C(Jin(x,y))}.
The feature vector
8 x,
y) ofa given pixel (x, y) used for
classification contains the values of the simplified image of
all the abundance maps and its local scale defined by Equa
tion (5), i.e.
In this section, we use the features defined by Equation (6)
to classify a hyperspectral image taken by the instrument RO
SIS (Reflective Optics System Imaging Spectrometer) over
the University
of
Pavia, Italy (see Figure 3(a)). The im
age (with a spatial resolution of 1.3m contains 340 x 610
pixels and 103 spectral bands covering visible and near in
frared light. The image is manually classified into 9 classes
4.
EXPERIMENT
(4)
x y _ L (k ,l )E ; (X,y )
k,
l
S(Ji(x, y))
The scale of this pixel E(x,y) is defined as the area of the
most contrasted shape divided by its perimeter, i.e.
Since the optical instruments always blur remote sensing im
ages, several shapes with very low contrasts can belong to
the same structure. In order to deal with the blur, the au
thors
of
[10] propose a geometrical criterion to cumulate the
contrasts of the shapes corresponding to one given structure.
The idea is that the difference
of
the areas
of
two succesive
shapes (for example f i and fi+l) corresponding to a same
structure is proportional to the perimeterof the smaller shape,
i.e.
S J
i+l) -
S J
i)
> P J
i), where> is a constant. It is
shown in [10] that this local scale corresponds very well to
the size of a structure and it is a very significant feature for
characterizing a structure in remote sensing images. Remark
that the most contrasted shapes extracted in an image form a
partition
of
this image. We can therefore have a simplified
image by defining the value of each pixel as the mean value
of
the gray levels
of
the pixels in its most contrasted shape,
i.e. the value of
x, y
in the simplified image is
where I( k, l) is the gray level
of
the pixel (k,l) in the orig
inal image. This value is more significant for characterizing
spectrally the pixel than its original gray level since it takes
into account the context pixels.
In order to extend the estimation of local scale to hyper
spectral images, the scales on all the abundance maps
of
the
p endmembers obtained by veA are first computed. There
fore, for each spatial position
x,
y in an hyperspectral im
age, there are p scale values. For the pixel
(x, y
on the
nth
abundancemap, we note
En(x, y)
its local scale calculated by
Equation (3), Ji
n
(x ,
y
the most contrasted shape extracted
,
by Equation (2), and In(x,y) its value
of
the simplified im
age defined by Equation (4). The simplest way to use the scale
features
ofa
pixel is to use all the En(x, y) values as features
for classification. However, the major drawback is that if an
Ji(x,y) = argmax{C Jj x , y)} (2)
J
-
8/9/2019 Ieee Mlsp 09 Luo Based
4/6
5.
CONCLUSION
(c)b)a)
Poin ts in Tra ining S
t
I
Thematic Colour
I
Trees 524
Asphalt
I I
Bitumen
375
Gravel
392
I I
(painted) metal sheets
265
Shadow
I
I
Self-Blocking Bricks
Meado
ws
I 1
Bare So il 53 2
Fig. 4. Definitions
of
the classes.
In this article, we have proposed to integrate geometrical fea
ture, the characteristic scales of structures, for the classifica
tion
of
hyperspectral images. In order to reduce the dimension
of
the data, we used a linear unmixing algorithm to extract the
the class
of
Meadows, which is not well defined inthe ground
truth.
It has to be remarked that by using only the spectral in
formation (i.e. the spectrum
of
each pixel), it is very difficult
to distinguish the class Asphalt and Bitumen, since they are
made by the same material. According to the ground truth
of
Figure 3(c), the only difference is that Asphalt is used for
the roads while Bitumen is used for the building roofs. How
ever, it can be seen from Figure 6(a) that the scale values
of
the pixels on the building roofs are very different from the
scales
of
the pixels on the roads. Therefore the classification
results
of
these two classes by using the feature vectors 8 are
much more accurate. This phenomenon can be illustrated by
Figure 6(d)-(t), where we have zoomed the classification re
sults obtained by using original data and the feature vectors
8 on a building roof made by bitumen.
can be seen that
the building roof and the road are classified as the same class
in Figure 6(e). In contrary, by adding the local scale feature,
these two structures are well separated.
Fig. 3. (a) Image
of
Pavia University (R-band 90, G-band 60,
B-band 40); (b) Training set; (d) Test set.
and the definitions of these classes are shown in Figure 4.
For classification purpose, the training set contains
3921
pix
els (see Figure 3(b)) while the test set contains 42776 pix
els (see Figure 3(c)). Since in this image, there are mainly
3 endmembers present (vegetation, bare soil and metal root),
we extract 3 endmembers and their abundance maps by us
ing VCA (see Section 2), which are shown in Figures 5(a)
(c). Afterwards we compute the simplified images
of
these
abundance maps by using Equation (4). The simplified im
ages are shown in Figure 5(d)-(t). According to Equation (5),
we estimate the local scale for each pixel of this image. The
scale map is shown in Figure 6(a). Therefore the feature vec
tor
8 x ,y
ofa pixel
(x ,y
defined by Equation (6) contains
only 4 values: 3 values
of
the simplified abundance maps
l l X,y),12(x,y), la(x,y)) and one scale value (E(x,y)).
We have classified the image by using the original hy
perspectral data (all
103
bands) and the feature vector 8 de
fined by Equation (6). Kernel methods are proved to be ef
fecient algorithms for hyperspectral image classification. [13]
Therefore, we use the Support Vector Machine (SVM) with
Gaussian kernel as classifier. The optimal scale parameter
of
Gaussian kernel is selected by 5-fold cross validation on the
training set.
The overall accuracies and the classification accuracies of
each class for both cases are shown in Table 1. In Figure 6(b)
and (c), the classification results obtained on the whole im
age is shown. In [5], the authors proposed to use PCA and
Kernel PCA (KPCA) to reduce the dimension
of
hyperspec
tral images. Extended Morphological Profiles (EMP) are then
extracted from the principle components obtained by PCA (3
principle components are extracted) or KPCA
(12
principle
components are extracted) as features for classifying the hy
perspectral image. In Table 1, we show the classification re
sults obtained in [5] on the same data set by using SVM with
Gaussian kernel.
It can be seen that by using the features proposed in our
article, the classification results improve considerably when
compared to the results obtained on original hyperspectral
data. Recall that the length
of
the feature vector 8 for each
pixel is only 4 which ismuch less than the number
of
spectral
bands
(103).
Moreover, the results obtained by using PCA
and EMP are very similar with the results obtained by the
features proposed, since we have previously mentioned that
the major information described by EMP on a pixel is its lo
cal scale. Indeed, the local scale
of
a pixel can be considered
as the width
of
the morphological filter by which the absolute
differential EMP value reaches its maximum. However, the
number of features extracted by PCA and EMP is 27, which
is much larger than the number
of
features in 8. The results
obtained by using KPCA and EMP are better than the results
obtained by using 8 and the results by using PCA and EMP.
However, the feature dimension, which is
108,
is even higher
than the dimension of original data
(103).
And the improve
ment of classification accuracies is mainly concentrated on
-
8/9/2019 Ieee Mlsp 09 Luo Based
5/6
Feature Original data Feature vector e
EMPp
C [ j
EMPK A[5j
Number
of
features 103
4
27 108
Overall accuracy 76 .01 91.54 92 .04 96 .55
Tree
98.59
94.52 99.22
99.35
Asphalt 78.16 96 .27 94 .60 96 .23
Bitumen 89 .02 99 .32 98.87 99 .10
Gravel 64 .55 84 .61 73.13 83 .66
(painted) metal sheets
99.47 99.55 99.55 99.48
Shadow
99.89 96.30 90.07
98.31
Self-Blocking Bricks 91.01
99.70 99.10 99.46
Meadows 64.23 85.80 88.79 97.58
Bare Soil 82 .72 96.56 95 .23 92 .88
Table 1. Classification results
(f)
e)
(b)
(d)
(a)
ig 6. (a) Scale map of the hyperspectral image, bright pixel
correspondsto big scale while dark pixel correspondsto small
scale; (b) Supervised classification results obtained on orig
inal data set; (c) Supervised classification results obtained
on the feature set obtained by Equation (6). (d)- (f ):
Zoom
around a bui lding on the: (d) original image; (e) classi fica
tion results of (b); (f) classification results of (c).
(f)
(c)
(e)
(b)
a)
(d)
ig 5. (a)-(c) Abundance maps of the 3 endmembers ex
tracted by VCA ; (d)-(f) the simplified images
of
these abun
dance maps.
-
8/9/2019 Ieee Mlsp 09 Luo Based
6/6
endmembers and their abundance maps contained in a hyper
spectral image. The abundance maps can be considered as a
compact representation of spectral information of this image,
since the number
of
endmembers contained in a hyperspec
tral image is much smaller than the numberof spectral bands.
With the help
of
these abundance maps, we extend the method
proposed in [10] to hyperspectral images to estimate the char
acteristic scales
of
the structures. The experiments show that
the use of the scale feature, the classification results improve
considerably, especially for the objects made by the same ma
terial but with different semantic meanings. By using the fea
tures proposed, the classification results are very similar to
the results obtained by using the methods based on PCA and
EMP, even though the number
of
features proposed is much
less.
Acknowledgement The authors would like to thank Pro
Paolo Gamba, University
of
Pavia, for providing the data and
the ground truth of the classification.
6. REFERENCES
[1] S. B. Serpico and L. Bruzzone, A new search algorithm
for feature selection in hyperspectralremote sensing im
ages, IEEE
Trans.
on Geoscience and Remote Sensing,
vol.
39,
no.
7,
pp.
1360 - 1367,
July
2001.
[2] B. Guo, S.R. Gunn, R. I. Damper, and J.D.B. Nelson,
Band selection for hyperspectral image classification
using mutual information, IEEE Geoscience and Re-
mote Sensing Letters, vol.
3,
no.
4,
pp.
522-526, 2006.
[3] D.A. Landgrebe, Signal Theory Methods in Multispec-
tral Remote Sensing, John Wiley and Sons, New Jersey,
2003.
[4] M. Fauvel, J. A. Benediktsson, J. Chanussot, and J. R.
Sveinsson, Spectral and spatial classification
of
hyper
spectral data using SVMs and morphological profile,
IEEE Transaction on Geoscience and Remote Sensing,
vol. 46, no. 11, November 2008.
[5] M. Fauvel, J. Chanussot, and J. A. Benediktsson, Ker
nel principal component analysis for the classification
of
hyperspectral remote-sensing data over urban areas,
EURASIP Journal on Advances in Signal Processing,
2009,
to appear.
[6] J.A. Palmason, J.A. Benediktsson, J.R. Sveinsson, and
J. Chanussot, Classification of hyperspectral data from
urban areas using morpholgical preprocessing and in
dependent component analysis, in IEEE IGARSS 05 -
International Geoscience and Remote Sensing Sympo-
sium, Seoul Korea, 2005, pp. 176-179.
[7] J. M. P.Nascimento and J. M. B. Dias, Vertex compo
nent analysis: A fast algorithm to unmix hyperspectral
data, IEEE Trans. Geoscience and Remote Sensing,
vol. 43, no. 4, pp. 898 - 910, April 2005.
[8] G. Poggi, G. Scarpa, and J. Zerubia, Supervised seg
mentation of remote sensing images based on a tree
structure MRF model, IEEE Trans. on Geoscience and
Remote Sensing, vol. 43, no. 8, pp. 1901-1911,2005.
[9] M. Pesaresi and J. A. Benediktsson, A new approach
for the morphological segmentation of high-resolution
satellite imagery, IEEE Trans. on Geoscience and Re-
mote Sensing, vol. 39, no. 2, pp. 309-320, Febrary 2001.
[10] B. Luo, J.-F. Aujol, and Y.Gousseau, Local scale mea
sure from the topographicmap and application to remote
sensing images,
SIAMMultiscale Modeling and Simu-
lation, to appear.
[11] V. Caselles, B. ColI, and J.-M. Morel, Topographic
maps and local contrast changes in natural images, Int.
J Compo
Vision,vol. 33, no. 1, pp. 5-27,1999.
[12]
P. Monasse and F. Guichard, Fast computation
of
a
contrast-invariant image representation, IEEE Trans.
on Image Processing,
vol.
9,
no.
5,
pp.
860-872,
may
2000.
[13] G. Camps-Valls and L. Bruzzone, Kernel-basedmeth
ods for hyperspectral image classification, IEEE Trans.
on Geoscience and Remote Sensing, vol. 43, no. 6, pp.
1351-1362,2005.